适用于可持续数据中心的理想 AI_ML 光互连解决方案.pdf

1、Ideal AI/ML Optical Interconnect Solutionfor Sustainable Data CentersIdeal AI/ML Optical Interconnect Solutionfor Sustainable Data CentersTaeyong KimCMO/LOPTICAL COMMUNICATION NETWORKSThe cost of GPU decreases due to improvements in computing hardware performance.Exponential increase in the availabi

2、lity and quality of data used to train AI models due to the rise of Internet and connectivity.More than 40%of households in the UK and US use ChatGPT.ChatGPT weekly active user 400 millionThe Rise of AIGPU Computation CostGrowth In the Use of Digital Technologies198119952022Source:IEA,Energy and AIW

3、orlds Most Powerful Supercomputershttps:/asc.llnl.gov/exascale/el-capitanEl CapitanSupercomputer Fugakuhttps:/www.r-ccs.riken.jp/en/fugaku Lawrence Livermore National Lab Peak Performance:2.79 Exaflops Peak Power:35 MW AMD EPYC 24C,AMD MI300A Slingshot-11 Interconnect RIKEN Center for Computational

4、Science,Japan Peak Performance:0.54 Exaflops Peak Power:29 MW Arm-based A64FX 48C Tofu Interconnect DFrontier DOE/SC/Oak Ridge National Laboratory Peak Performance:2.06 Exaflops Peak Power:24 MW AMD EPYC 64C,AMD MI250X Slingshot-11 InterconnectEl Capitan,Frontier and Fugaku are with 11 million,9 mil

5、lion and 7 million cores,respectively.The peak power consumption ranges from 24 to 35 MW.The three supercomputers are liquid cooled.100,000 active optical cables over 900 km optical fibers are used in Supercomputer Fugaku.https:/www.oclf.ornl.gov/frontierOutlook for Electricity ConsumptionFour times

6、 greater electricity consumption may be needed in 5 years from today.The increase in the capacity of accelerated servers is one of the main drivers for electricity consumption.Energy efficient operation of data centers are needed.Estimated Datacenter Electricity Demand by AIBreakdown of Electricity